ASPECT EXTRACTION IN E-COMMERCE USING LATENT DIRICHLET ALLOCATION (LDA) WITH TERM FREQUENCY-INVERSE DOCUMENT FREQUENCY (TF-IDF)

Abstract

Social media is a common thing that people use. Posts or comments found on social media describe someone’s feelings and opinions so there have to be important topics that can be extracted from social media. In the e-commerce field, topic is an interesting thing to know because it can describes people’s opinion towards a product. However, the large number of social media users is currently making the process of finding topics from social media difficult, so computer assistance is needed. One method that can be used is Latent Dirichlet Allocation (LDA). LDA is a good method for extracting topics, but the drawback is that sometimes the topics are incomprehensible. To cover up the drawback, TF-IDF feature selection method is used so that less important words can be skipped so LDA can generate a better topic. The best hyperparameter values ​​obtained were 10 iterations, 10 topics, α and β values consecutively 0,1 and 0,01. The best feature selection percentile value is 90. This value is used to find the threshold that can be used as the lower limit of the TF-IDF value of each word so that the word with greater TF-IDF value can be used as feature.

Keywords

• Aspect Extraction
• Latent Dirichlet Allocation
• Perplexity
• Term Frequency - Inverse Document Frequency
• Topic Modelling